Synergistic use of spectral features of leaf nitrogen and physiological indices improves the estimation accuracy of nitrogen concentration in rapeseed

Abstract

ABSTRACT Estimation of nitrogen content in crop leaves using hyperspectral techniques encounters the challenge of limited model stability. In this study, a method for the construction of nitrogen estimation model of oilseed rape (Brassica napus L.) was proposed, which combines the spectral features of nitrogen and nitrogen-containing physiological components. Based on the pot experiment, a total of 320 sets of canopy spectra were collected in the five growth stages using hyperspectral technology, and the concentration of leaf nitrogen (LNC), malondialdehyde (MDA), proline, soluble sugar (SS), and water were determined in indoor chemical experiments. Original spectra was preprocessed by five kinds of spectral conversions, and the random frog (RF) method was used to screen the spectral features of nitrogen and nitrogen-containing physiological components that were used to construct the partial least square regression (PLSR) and support vector machine (SVM) models. The results showed that the selection of spectral features of LNC and nitrogen-containing physiological components based on RF greatly reduced the collinearity and redundant information among bands. The accuracy of the models based on the spectral features of LNC and nitrogen-containing physiological components were higher than that of the model based on the spectral features of single LNC. Further interannual test showed that the prediction accuracy of the N + MDA + SS model was higher than that of other models, especially at the rapid growth stage, with R2 of 0.871. This study provides a new idea for improving the accuracy and stability of crop nitrogen estimation model.